Root-cause analysis (RCA) in large-scale distributed systems is notoriously difficult because a single symptom can cut across service, data, configuration, compute, and organizational boundaries. Responders often have to piece together incomplete evidence under time pressure, making it hard to pinpoint the true source of a failure.
The arXiv paper introduces Diviner, an agentic RCA system that incorporates human-governed adaptation. Rather than fully automating the process, Diviner appears to let agents work through evidence while humans retain control over how the analysis adapts. This design aims to combine the scalability of automated reasoning with the judgment of experienced responders.
The abstract frames the problem as one of assembling incomplete evidence across multiple domains. By explicitly addressing that challenge, Diviner targets a practical pain point for operators of large infrastructure. The paper's contribution is the system itself, though details of its implementation and evaluation are not covered in the abstract.